CAJ | 학술논문

基于单图像的3D重构因其先天性约束不足和潜在的巨大价值成为计算机视觉领域的研究热点，被广泛应用于航空航天、机械制造、医疗、考古、地质、犯罪现场复原、建筑设计、城市规划等领域。针对图像中几何元素的共平面性可提供景深信息的特点，提出一种基于交叉曲线共平面性约束的3D重构方案，即对于不平行于投影方向的某一个平面，根据其所含曲线与另一平面中某曲线的交叉构型构造一个线性系统，当一组这样的交叉曲线位于拟求解表面时，可获得精确解。对于含噪系统，要求测得的交点远离平坦面，增加新的约束条件，定义表面的平坦度度量模型，使用SVD法获得极小化线性系统的代数误差的逼近解。由于利用正投影和透视投影的等价性，可将透视投影转化为正投影，从而将这两种投影下的3D重构规划一个框架中。实验表明，这种方法大大提高了3D重构的健壮性，对噪声的敏感性小，可适用于完全未标定结构光等真实场景。
기우단도상적3D중구인기선천성약속불족화잠재적거대개치성위계산궤시각영역적연구열점，피엄범응용우항공항천、궤계제조、의료、고고、지질、범죄현장복원、건축설계、성시규화등영역。침대도상중궤하원소적공평면성가제공경심신식적특점，제출일충기우교차곡선공평면성약속적3D중구방안，즉대우불평행우투영방향적모일개평면，근거기소함곡선여령일평면중모곡선적교차구형구조일개선성계통，당일조저양적교차곡선위우의구해표면시，가획득정학해。대우함조계통，요구측득적교점원리평탄면，증가신적약속조건，정의표면적평탄도도량모형，사용SVD법획득겁소화선성계통적대수오차적핍근해。유우이용정투영화투시투영적등개성，가장투시투영전화위정투영，종이장저량충투영하적3D중구규화일개광가중。실험표명，저충방법대대제고료3D중구적건장성，대조성적민감성소，가괄용우완전미표정결구광등진실장경。
3D Reconstruction based on a single image is the research hotspot of computer vision because of its natural under-constrained and huge potential worthiness, which have been applied to aviation, mechanism, archaeology, geology, recovering the crime scene, architecture and city planning, etc. By the fact that co-plane of geometric item in an image may be provide the information about depth of field, this article suggested a complete new idea for 3D reconstruction based on co-plane constraint about intersection curves. For the planes which do not contain the projection direction, one formulated a linear system by the configuration of across curves between planes.When such a set on curves lied in a solution surface, one can get a accurate solution space. For any noisy systems, a key is that data points would escape from flat and nearly-flat planes, hence to add a new constraint conditions , define a measure of surface flatness and use SVD to obtain a approximate solution that minimizes the algebraic error of the linear system. On the other hand, since the equivalent between orthographic projection and perspective projection, one can resolve perspective projection via orthographic projection, and thereby could layout a unitive 3D reconstruction formula. Experimentation demonstrated that aforementioned methods advanced the robustness of 3D reconstruction, and have less sensitive to noise, would be suitable for real scene of completeness uncalibrated structured light.